Semantic Segmentation of Satellite Images for Water Body Detection

Abstract

Semantic Segmentation is a technique in Computer vision which is used to label different pixel classes present in imagery. When talking about Satellite images, Semantic Segmentation can be used to segment water bodies, forest area, buildings, agriculture lands, etc. In this paper, we have focused on Semantic Segmentation of Satellite images for Water Body Detection. We have used a predefined architecture U-Net as the segmentation model to detect and segment water bodies present in satellite imagery. U-Net is a better choice over other models because of its various advantages like less training data required, low inference time and fast learning. We have used sentinel-2 dataset for training and validation of the model. The images are RGB images and they are stored in jpeg format. The dataset has 2000 + images, however, we have tried to train and validate our model on minimum number of images so as to get an optimal model in less amount of data. We succeeded in achieving good results in less data. The final result is the segmented mask of a satellite image where white region is water body and black region is land area.